Unbalance data set-oriented mine ventilation system resistance change type fault diagnosis method based on Bayesian optimization depth Q network

Abstract

The invention belongs to the technical field of mine ventilation, and particularly relates to a mine ventilation system resistance change type fault diagnosis method based on a Bayesian optimization depth Q network for an imbalance data set, which comprises the following steps: firstly, generating a simulation data fault sample set about resistance change type fault-air volume through network calculation of the mine ventilation system; and then, a fault diagnosis model WAFs-BODQN based on a Bayesian optimization (BO) deep Q network (DQN) is constructed by taking an air volume parameter of a mine ventilation system as an environment state space of the model and taking a branch number as an action of an intelligent agent, and in order to solve the problem that the intelligent model is poor in performance capability during data imbalance, a fault diagnosis model WAFs-BODQN is established. A reinforcement learning reward function KMeans-SDF combining a KMeans + + algorithm and a spatial distance function (SDF) is designed, the phenomenon of eccentricity of majority of samples caused by an intelligent model can be effectively avoided, accurate prediction of fault positions is achieved, a general system structure for resistance-variable fault diagnosis of the mine ventilation system is expected to be established according to the method, and the method is suitable for popularization and application. Theoretical guidance and technical support are provided for intelligent ventilation of the mine.

Description

Technical Field

[0001] The present invention relates to the technical field of mine ventilation, and in particular to a mine ventilation system resistance-variable fault diagnosis method based on a Bayesian optimized deep Q network for an unbalanced data set. Background Art

[0002] The mine ventilation system is responsible for providing fresh air flow underground, exhausting polluted air and diluting toxic and harmful gases. It plays an indispensable role in ensuring mine safety, disaster prevention, disaster reduction and disaster relief. With the increase of mining depth, the ground stress of the tunnel increases, the stability of the surrounding rock decreases, and the tunnel deformation, caving, slab collapse, roof fall and other resistance type faults (WAFs) occur. After the resistance type fault occurs, it will cause changes in the air volume of the local or even the entire mine, resulting in the accumulation of toxic gases and dust, and even fire and explosion accidents in severe cases. Therefore, it is necessary to locate the location of the resistance type fault in time. In recent years, driven by the development of artificial intelligence, mine ventilation system fault diagnosis has entered the intelligent era. The current intelligent diagnosis model is mainly based on traditional supervised learning algorithms such as support vector machine, multi-layer perceptron, decision tree, K nearest neighbor algorithm and unsupervised learning algorithms such as genetic algorithm and evolutionary strategy, which have achieved good fault diagnosis effect. However, unsupervised learning algorithm modeling is difficult and time-consuming. Supervised learning algorithm passively learns the nonlinear mapping relationship between input and output, lacks autonomous decision-making ability and poor generalization, and it is difficult to establish a general architecture for the diagnosis of resistance-change type faults in mine ventilation systems. In addition, the research on resistance-change type fault diagnosis is carried out under the ideal condition of balanced data sets. However, in the actual production process of mines, due to the different tunnel support and environment, the probability of resistance-change type faults in each tunnel is also different, resulting in different numbers of fault samples in different tunnels. Therefore, the situation of sample imbalance is an objective existence. In the case of sample imbalance, the model will learn the prior knowledge of sample proportion, which will lead to the model's bias towards the category with more sample data when making predictions, affecting the more essential features of model learning, and thus affecting the robustness of the model. Therefore, this paper proposes a mine ventilation system resistance fault diagnosis method WAFs-BODQN based on (BO) deep Q network algorithm (DQN) for imbalanced data sets. At the same time, in order to avoid the interference of sample imbalance on the learning effect of intelligent model, a reinforcement learning reward function Kmeans-SDF combining KMeans++ clustering algorithm and spatial distance function (SDF) is designed. Based on this method, it is expected to establish a general architecture for mine ventilation system fault diagnosis, provide theoretical guidance and technical support for the realization of mine intelligent ventilation, and have important theoretical significance and application value. Summary of the invention

[0003] In view of the deficiencies in the prior art, a method for diagnosing resistive-type faults of mine ventilation systems based on a Bayesian optimized deep Q network for unbalanced data sets is provided. The method is characterized in that a resistive-type fault diagnosis model for mine ventilation systems based on a Bayesian optimized deep Q network is constructed, and a reward function combining the Kmeans++ clustering algorithm and the spatial distance function is set, which solves the problem of the model's poor performance when the samples are unbalanced, resulting in the "eccentricity" phenomenon of the majority class samples, and achieves accurate prediction of the fault location. According to this method, it is expected to establish a general architecture for the diagnosis of resistive-type faults of mine ventilation systems, and provide theoretical guidance and technical support for the realization of intelligent ventilation in mines. In order to solve the above-mentioned technical problems, the present invention proposes a method for diagnosing resistive-type faults of mine ventilation systems based on a Bayesian optimized deep Q network for unbalanced data sets, comprising the following steps: Step 1: prepare the topological relationship data of the mine ventilation network, the topological relationship data includes the original wind resistance and air volume of each branch of the mine ventilation network, and the fan characteristic curve of the ventilator. The fan characteristic curve equation is specifically expressed as: (1) In the formula, Indicates the wind pressure of the fan. Indicates the air volume of the ventilator. Indicates the fan wind pressure characteristic coefficient; Step 2: Set the imbalance ratio, solve the mine ventilation network, obtain the wind resistance and air volume of each branch of the mine ventilation network after the fault, and generate an imbalance fault sample set about "resistance change fault-air volume", the fault sample set includes the fault number and fault volume of each branch of the mine ventilation network, and the wind resistance value and air volume of each branch of the mine ventilation network after the fault occurs; Step 3: According to the fault sample set, a Markov decision process for determining the resistance-variable fault diagnosis of the mine ventilation system is constructed, the structure and reward function of the Q network are designed, and a fault diagnosis model WAFs-BODQN based on the Bayesian optimization deep Q network is constructed.

[0004] The step 2 sets the imbalance ratio, and obtains the wind resistance and air volume of each branch of the mine ventilation network after the fault through the mine ventilation network solution, and generates an imbalance fault sample set about "resistance change fault-air volume". The fault sample set includes the fault number and fault volume of each branch of the mine ventilation network, and the wind resistance value and air volume of each branch of the mine ventilation network after the fault occurs, which is specifically expressed as: Step 2.1: Set the imbalance ratio. Sample imbalance occurs in classification problems. In the field of resistance-variable fault diagnosis of mine ventilation systems, sample imbalance means that the probability of failure in each lane is different, resulting in a different number of fault samples in each branch. The imbalance ratio (IR) is defined as: (2) In the formula, is the number of majority class samples, is the number of minority class samples; Sample imbalance is divided into three levels: (1) mild imbalance: IR value range is [2.5, 5], (2) moderate imbalance: IR value range is [5, 100], (3) extreme imbalance: IR>100. In order to obtain sample sets with different imbalance levels, different sample imbalance ratios are set respectively; Step 2.2: Define the original air volume matrix of all branches of the mine ventilation network as Q, the original wind resistance matrix as R, and the original wind pressure matrix as , recorded as: (3) And meet: (4) (5) In the formula, Represents the original air volume of the nth network branch, represents the original wind resistance of the nth network branch, h n represents the original wind pressure of the nth network branch, represents the complete incidence matrix of the mine ventilation network, represents the original air volume of the jth network branch, represents the i-th row and j-th column element of the complete incidence matrix of the mine ventilation network, m represents the number of nodes in the mine ventilation network, represents the circuit matrix of the mine ventilation network, represents the i-th row and j-th column element of the circuit matrix of the mine ventilation network, s represents the total number of circuits in the mine ventilation network, n represents the number of branches in the mine ventilation network, Represents the original wind pressure matrix The transpose of Represents the additional resistance matrix of all circuits in the mine ventilation network, Represents the additional resistance matrix of all circuits in the mine ventilation network The transpose of Represents the original wind pressure matrix of the mine ventilation network The jth column of Representing the circuit additional resistance matrix of the mine ventilation network The i-th row of Step 2.3: Hypothesis Branching A resistance change fault occurs, and the fault quantity is , using formula (6) to calculate the branch The branch wind resistance after the failure occurs, then the wind resistance matrix composed of all branches of the mine ventilation network after the failure occurs Expressed as , , (6) In the formula, Indicates branch e i Branch windage after a fault occurs, Represents a branch The original wind resistance, Indicates branch e i The number of faults that occurred, n represents the number of all branches of the mine ventilation network; Step 2.4: Solve the air volume of the mine ventilation network after the failure through network solution. The air volume matrix composed of the air volume of all branches of the mine ventilation network after the failure is Expressed as , the network solution is expressed as: (7) In the formula, Indicates the air volume of the mine ventilation network after the failure, represents the wind resistance of the mine ventilation network after the fault, Represents the function of wind resistance and wind volume; Step 2.5: The calculated wind resistance matrix of the mine ventilation network after the failure , air volume matrix , the fault branch number each time a fault occurs , Fault quantity Recorded in the fault sample set.

[0005] The step 3 described above determines the Markov decision process for the resistance-variable fault diagnosis of the mine ventilation system, designs the structure and reward function of the Q network, and constructs the fault diagnosis model WAFs-BODQN based on the Bayesian optimization deep Q network, which is specifically expressed as follows: Step 3.1: Markov decision process for fault diagnosis of mine ventilation system Randomly select l samples D={(Q1,e1), (Q2,e2), …, (Q i ,e i ).., (Q l ,e l )} constitutes l guessing questions, and then the agent guesses in the order of samples in D, where Q i is the air volume data corresponding to the sample, e i is the corresponding label, and finally the resistance-variable fault diagnosis problem of the mine ventilation system is transformed into a classification Markov decision process (CMDP), which is expressed as: ; In the formula, S represents the environmental state set, which depends on the training sample. In this study, it is the air volume set of each branch of the mine after the resistance type fault occurs. The first training sample As the initial state s1, state s t Corresponding sample x t , n is the number of branches; A represents the action set, that is, the set of fault branch numbers ; R is the reward function, which comes from the feedback of the environment and is used to measure the pros and cons of the agent's execution of a certain action and guide the agent to learn the best strategy; P is the state transition probability matrix, which represents the rule of transitioning from the current state to the next state after executing the action ,In this study, P is determined and transferred according to the order of the samples in the training samples.

[0006] Step 3.2: WAFs-BODQN model structure design The WAFs-BODQN model contains two networks with the same structure but different parameters: Eval-Net and Target-Net. Eval-Net is used to update the current Q value, and Target-Net is used to calculate the fixed target Q value. One-dimensional convolutional neural network (1DCNN) is used as Eval-Net and Target-Net of WAFs-BODQN, and BO is used to optimize the parameters of 1DCNN. The whole process is expressed as follows: (8) In the formula, is the activation function used in the output layer, pool represents the pooling operation, Represents the activation function used by the activation layer, is the set of feature maps output by the l-1th layer, is the jth output feature map of the l-1th layer, * is the convolution operation, is the i-th convolution kernel of the l-th layer, is the i-th bias of the l-th layer; Using BO to adjust the hyperparameters of 1DCNN To optimize, the probability distribution of the optimization process is established using the Bayesian formula, which is expressed as formula (9): (9) In the formula, is the posterior probability distribution of y, is the likelihood distribution of y, represents the marginal likelihood distribution, is the prior probability distribution of y, is the objective function value, is the objective function, is the observation noise, is the set of historical observations, is the decision vector, i.e., the hyperparameter of CNN.

[0007] Step 3.3: Reward function design Based on the KMeans++ clustering algorithm and the spatial distance function, the reward function is quantified by the Euclidean distance between samples. First, a sample is randomly selected as the cluster center. , and then calculate each sample To cluster center Minimum distance: (10) Select the sample point with the largest distance as the next cluster center , repeat the above operation to finally get k clusters. Then, calculate the Euclidean distance between the samples in the same cluster and the cluster center and perform normalization: (11) (12) Set the reward value of the sample corresponding to the center of the majority class cluster to 1, and the reward value of the remaining samples to 1. As a bias, the reward value of the sample corresponding to the center of the minority cluster is set to the imbalance ratio IR, and the remaining samples are As a bias, the reward function is: (13) The reward function R3 fully considers the imbalance and diversity of samples with the help of class spacing and intra-class spacing, enabling the agent to achieve better learning results.

[0008] Step 3.4: Test the resistance-variable fault diagnosis method of the mine ventilation system based on the Bayesian optimized deep Q network constructed for the unbalanced data set, and verify the accuracy of the model diagnosis by comparing whether the fault branch number predicted by the model is the same as the actual fault branch number in the fault sample set.

[0009] The beneficial effects of the present invention are: The present invention provides a method for diagnosing resistive faults of mine ventilation systems based on a Bayesian optimized deep Q network for unbalanced data sets. The method realizes an end-to-end architecture from the state space of the mine ventilation system to the fault location, is more intelligent and generalized, and can avoid the interference of sample imbalance on the fault diagnosis results, thereby achieving accurate prediction of the fault location. Based on this method, it is expected to establish a general architecture for diagnosing resistive faults of mine ventilation systems. BRIEF DESCRIPTION OF THE DRAWINGS

[0010] Figure 1This is a flow chart of a method for diagnosing resistance-variable faults in mine ventilation systems based on Bayesian optimization deep Q-network for unbalanced data sets.

[0011] Figure 2 This is a T-type mine ventilation network diagram.

[0012] Figure 3 This is the structure diagram of BO-1DCNN.

[0013] Figure 4 This is the clustering result of PCA-KMeans++ algorithm for different imbalanced samples.

[0014] Figure 5 This is the fault diagnosis result diagram of the T-type mine ventilation network. DETAILED DESCRIPTION

[0015] The invention is further described below in conjunction with the accompanying drawings and specific implementation examples: A method for diagnosing resistance-variable faults in a mine ventilation system based on a Bayesian optimized deep Q network, the process of which is as follows: Figure 1 As shown, the specific steps include: Step 1: prepare the topological relationship data of the mine ventilation network, the topological relationship data includes the original wind resistance and air volume of each branch of the mine ventilation network, and the fan characteristic curve of the ventilator. The fan characteristic curve equation is specifically expressed as: (1)

[0016] In the formula, Indicates the wind pressure of the fan. Indicates the air volume of the ventilator. Indicates the fan wind pressure characteristic coefficient;

[0017] like Figure 2 Mine ventilation network shown It is a single-source single-sink network with three diagonal branches. is the branch set, the number of branches , is the node set, the number of nodes Among them, the main fan working point air volume is 87 m 3 / s, the wind pressure is 2281.77 Pa, and the fan characteristic curve is The original data and initial state parameters of the network are shown in Table 1.

[0018] Table 1 Original data of ventilation network G

[0019] Branch number <![CDATA[Air resistance r (N·s 2 ·m -8 )]]> <![CDATA[Air volume q (m 3 ·s -1 )]]> Wind pressure h(Pa) Node Number Node pressure energy P(Pa) 1 0.05 86.99 408.7 1 0 2 0.14 44.58 289.6 2 -408.7 3 0.13 42.41 241.2 3 -698.3 4 0.53 42.47 962.7 4 -649.9 5 6.12 2.11 27.4 5 -725.7 6 1.38 7.40 75.8 6 -876.5 7 1.66 9.51 150.7 7 -1661.1 8 0.18 35.00 226.6 8 -0.002 9 0.39 44.52 784.5 10 0.08 86.99 620.6

[0020] Step 2: Set the imbalance ratio, solve the mine ventilation network, obtain the wind resistance and air volume of each branch of the mine ventilation network after the fault, and generate an imbalance fault sample set about "resistance change fault-air volume", the fault sample set includes the fault number and fault volume of each branch of the mine ventilation network, and the wind resistance value and air volume of each branch of the mine ventilation network after the fault occurs;

[0021] Step 2.1: Set the imbalance ratio. Sample imbalance occurs in classification problems. In the field of resistance-variable fault diagnosis of mine ventilation systems, sample imbalance means that the probability of failure in each lane is different, resulting in a different number of fault samples in each branch. The imbalance ratio (IR) is defined as:

[0022] (2)

[0023] In the formula, is the number of majority class samples, is the number of minority class samples;

[0024] Sample imbalance is divided into three levels: (1) mild imbalance: IR value range is [2.5, 5], (2) moderate imbalance: IR value range is [5, 100], (3) extreme imbalance: IR>100. In order to obtain sample sets with different imbalance levels, different sample imbalance ratios are set.

[0025] Step 2.2: Define the original air volume matrix of all branches of the mine ventilation network as Q, the original wind resistance matrix as R, and the original wind pressure matrix as , recorded as:

[0026] (3)

[0027] And meet:

[0028] (4)

[0029] (5)

[0030] In the formula, Represents the original air volume of the nth network branch, represents the original wind resistance of the nth network branch, h n represents the original wind pressure of the nth network branch, represents the complete incidence matrix of the mine ventilation network, represents the original air volume of the jth network branch, represents the i-th row and j-th column element of the complete incidence matrix of the mine ventilation network, m represents the number of nodes in the mine ventilation network, represents the circuit matrix of the mine ventilation network, represents the i-th row and j-th column element of the circuit matrix of the mine ventilation network, s represents the total number of circuits in the mine ventilation network, n represents the number of branches in the mine ventilation network, Represents the original wind pressure matrix The transpose of Represents the additional resistance matrix of all circuits in the mine ventilation network, Represents the additional resistance matrix of all circuits in the mine ventilation network The transpose of Represents the original wind pressure matrix of the mine ventilation network The jth column of Representing the circuit additional resistance matrix of the mine ventilation network The i-th row of

[0031] Step 2.3: Hypothesis Branching A resistance change fault occurs, and the fault quantity is , using formula (6) to calculate the branch The branch wind resistance after the failure occurs, then the wind resistance matrix composed of all branches of the mine ventilation network after the failure occurs Expressed as ,

[0032] , (6)

[0033] In the formula, Indicates branch e i Branch windage after a fault occurs, Represents a branch The original wind resistance, Indicates branch e i The number of faults that occurred, n represents the number of all branches of the mine ventilation network;

[0034] Step 2.4: Solve the air volume of the mine ventilation network after the failure through network solution. The air volume matrix composed of the air volume of all branches of the mine ventilation network after the failure is Expressed as , the network solution is expressed as:

[0035] (7)

[0036] In the formula, Indicates the air volume of the mine ventilation network after the failure, represents the wind resistance of the mine ventilation network after the fault, Represents the function of wind resistance and wind volume;

[0037] Step 2.5: The calculated wind resistance matrix of the mine ventilation network after the failure , air volume matrix , the fault branch number each time a fault occurs , Fault quantity Recorded in the fault sample set.

[0038] Step 3: According to the fault sample set, a Markov decision process for determining the resistance-variable fault diagnosis of the mine ventilation system is constructed, the structure and reward function of the Q network are designed, and a fault diagnosis model WAFs-BODQN based on the Bayesian optimization deep Q network is constructed;

[0039] Step 3.1: Markov decision process for fault diagnosis of mine ventilation system

[0040] Randomly select l samples D={(Q1,e1), (Q2,e2), …, (Q i ,e i ).., (Q l ,e l )} constitutes l guessing questions, and then the agent guesses in the order of samples in D, where Q i is the air volume data corresponding to the sample, e i is the corresponding label, and finally the resistance-variable fault diagnosis problem of the mine ventilation system is transformed into a classification Markov decision process (CMDP), which is expressed as: .

[0041] In the formula, S represents the environmental state set, which depends on the training sample. In this study, it is the air volume set of each branch of the mine after the resistance type fault occurs. The first training sample As the initial state s1, state s t Corresponding sample x t , n is the number of branches; A represents the action set, that is, the set of fault branch numbers ; R is the reward function, which comes from the feedback of the environment and is used to measure the pros and cons of the agent's execution of a certain action and guide the agent to learn the best strategy; P is the state transition probability matrix, which represents the rule of transitioning from the current state to the next state after executing the action ,In this study, P is determined and transferred according to the order of the samples in the training samples.

[0042] Step 3.2: WAFs-BODQN model structure design

[0043] The WAFs-BODQN model contains two networks with the same structure but different parameters: Eval-Net and Target-Net. Eval-Net is used to update the current Q value, and Target-Net is used to calculate the fixed target Q value. One-dimensional convolutional neural network (1DCNN) is used as Eval-Net and Target-Net of WAFs-BODQN, and BO is used to optimize the parameters of 1DCNN. The whole process is expressed as follows:

[0044] (8)

[0045] In the formula, is the activation function used in the output layer, pool represents the pooling operation, Represents the activation function used by the activation layer, is the set of feature maps output by the l-1th layer, is the jth output feature map of the l-1th layer, * is the convolution operation, is the i-th convolution kernel of the l-th layer, is the i-th bias of the l-th layer;

[0046] Using BO to adjust the hyperparameters of 1DCNN To optimize, the probability distribution of the optimization process is established using the Bayesian formula, which is expressed as formula (9):

[0047] (9)

[0048] In the formula, is the posterior probability distribution of y, is the likelihood distribution of y, represents the marginal likelihood distribution, is the prior probability distribution of y, is the objective function value, is the objective function, is the observation noise, is the set of historical observations, is the decision vector, i.e., the hyperparameter of CNN.

[0049] Step 3.4: Reward function design

[0050] Based on the KMeans++ clustering algorithm and the spatial distance function, the reward function is quantified by the Euclidean distance between samples. First, a sample is randomly selected as the cluster center. , and then calculate each sample To cluster center Minimum distance:

[0051] (10)

[0052] Select the sample point with the largest distance as the next cluster center , repeat the above operation to finally get k clusters. Then, calculate the Euclidean distance between the samples in the same cluster and the cluster center and perform normalization:

[0053] (11)

[0054] (12)

[0055] Set the reward value of the sample corresponding to the center of the majority class cluster to 1, and the reward value of the remaining samples to 1. As a bias, the reward value of the sample corresponding to the center of the minority cluster is set to the imbalance ratio IR, and the remaining samples are As a bias, the reward function is:

[0056] (13)

[0057] The reward function R3 fully considers the imbalance and diversity of samples with the help of class spacing and intra-class spacing, enabling the agent to achieve better learning results.

[0058] Step 3.5: Test the resistance-variable fault diagnosis method of the mine ventilation system based on the Bayesian optimized deep Q network constructed for the unbalanced data set, and verify the accuracy of the model diagnosis by comparing whether the fault branch number predicted by the model is the same as the actual fault branch number in the fault sample set.

[0059] According to the specific steps given in step 2 above, the sample imbalance ratios are set to IR=3, 8, and 103 respectively, and three groups of sample sets with different imbalance degrees are obtained, namely UB1, UB2, and UB3, which correspond to the slightly imbalanced sample set, the moderately imbalanced sample set, and the extremely imbalanced sample set, respectively. The number of samples corresponding to each branch in each sample set and the total number of samples in the sample set are shown in Table 2; According to the specific steps described in step 2, a test sample set is generated, which contains 80 samples. The test sample set is shown in Table 3;

[0060] Table 2 Sample number configuration of each category of unbalanced training samples

[0061] Imbalanced training sample set number IR e2 e3 e4 e5 e5 e6 e7 e8 total UB1 3 792 792 792 792 792 792 792 2376 7920 UB2 8 528 528 528 528 528 528 528 4224 7920 UB3 103 72 72 72 72 72 72 72 7416 7920

[0062] Table 3 Test samples

[0063] Sample No. q1 q2 q3 q4 q5 q6 q7 q8 q9 q10 r eei 1 79.55 25.25 54.31 32.79 -7.55 13.46 5.91 40.85 46.76 79.55 1.42 2 2 88.06 47.89 40.17 43.95 3.95 6.03 9.97 34.14 44.11 88.06 -0.07 2 3 84.07 35.33 48.75 39.01 -3.68 11.10 7.42 37.65 45.07 84.07 0.31 2 4 82.38 31.25 51.12 36.70 -5.45 12.18 6.73 38.95 45.68 82.38 0.61 2 5 79.81 25.76 54.05 33.15 -7.39 13.36 5.97 40.69 46.66 79.81 1.32 2 6 80.51 27.21 53.30 34.13 -6.91 13.07 6.16 40.23 46.39 80.51 1.08 2 … … … … … … … … … … … … … 79 85.64 44.88 40.76 43.75 1.14 7.54 8.68 33.22 41.90 85.64 0.11 9 80 89.35 43.70 45.65 40.17 3.52 7.35 10.87 38.30 49.17 89.35 -0.15 9

[0064] According to the specific steps given in step 3 above, the Markov decision process of the ventilation network G is obtained , the environment state space , the action space is ; The Bayesian optimized one-dimensional convolutional neural network structure is as follows Figure 3 As shown in Figure 2, the clustering results of the KMeans++ clustering algorithm for the imbalanced sample sets UB1, UB2, and UB3 are as follows: Figure 4 As shown; Finally, the WAFs-BODQN model is trained using the imbalanced sample sets UB1, UB2, and UB3, and fault diagnosis is performed on the test sample set. The diagnosis results are shown in Figure 5 As shown in the figure, the fault location diagnosis accuracy reached 98.75%, 97.50%, and 95% respectively, achieving accurate prediction of the fault location.

Claims

1. A method for diagnosing resistance-variable faults in mine ventilation systems based on Bayesian optimization deep Q-network for unbalanced data sets, characterized in that: The following steps are involved: Step 1: prepare the topological relationship data of the mine ventilation network, the topological relationship data includes the original wind resistance and air volume of each branch of the mine ventilation network, and the fan characteristic curve of the ventilator. The fan characteristic curve equation is specifically expressed as: H τ =t0+t1Q τ +t2Q τ 2 (1) In the formula, H τ Indicates the wind pressure of the fan, Q τ represents the air volume of the fan, t0, t1, t2 represent the fan pressure characteristic coefficients; Step 2: Set the imbalance ratio, solve the mine ventilation network, obtain the wind resistance and air volume of each branch of the mine ventilation network after the fault, and generate an imbalance fault sample set about "resistance change fault-air volume", which includes the fault number and fault volume of each branch of the mine ventilation network, and the wind resistance value and air volume of each branch of the mine ventilation network after the fault occurs; Step 3: According to the fault sample set, a Markov decision process for determining the resistance-variable fault diagnosis of the mine ventilation system is constructed, the structure and reward function of the Q network are designed, and a fault diagnosis model WAFs-BODQN based on the Bayesian optimization deep Q network is constructed.

2. According to claim 1, a method for diagnosing resistance-variable faults in a mine ventilation system based on a Bayesian optimized deep Q network is characterized in that: The step 2 is specifically described as follows: Step 2.1: Set the imbalance ratio. Sample imbalance occurs in classification problems. In the field of resistance-variable fault diagnosis of mine ventilation systems, sample imbalance means that the probability of failure in each lane is different, resulting in a different number of fault samples in each branch. The imbalance ratio (IR) is defined as: Where n M is the number of majority class samples, n m is the number of minority class samples; Sample imbalance is divided into three levels: (1) mild imbalance: IR value range is [2.5, 5], (2) moderate imbalance: IR value range is [5, 100], (3) extreme imbalance: IR>100. In order to obtain sample sets with different imbalance levels, different sample imbalance ratios are set respectively; Step 2.2: Define the original air volume matrix of all branches of the mine ventilation network as Q, the original wind resistance matrix as R, and the original wind pressure matrix as H, recorded as: And meet: In the formula, q n represents the original air volume of the nth network branch, r n represents the original wind resistance of the nth network branch, h n represents the original wind pressure of the nth network branch, B = (b ij ) m×n represents the complete incidence matrix of the mine ventilation network, q j represents the original air volume of the jth network branch, b ij represents the i-th row and j-th column element of the complete incidence matrix of the mine ventilation network, m represents the number of nodes in the mine ventilation network, C = (c ij ) s×n represents the circuit matrix of the mine ventilation network, c ij represents the i-th row and j-th column element of the circuit matrix of the mine ventilation network, s represents the total number of circuits in the mine ventilation network, n represents the number of branches in the mine ventilation network, H T represents the transpose of the original wind pressure matrix H, H η represents the additional resistance matrix of all circuits in the mine ventilation network, H η T The additional resistance matrix H represents all circuits of the mine ventilation network η The transpose of h j represents the jth column of the original wind pressure matrix H of the mine ventilation network, h ηi The loop additional resistance matrix H represents the mine ventilation network η The i-th row of Step 2.3: Assume branch e i A resistance change fault occurs, and the fault quantity is Δr i , use formula (5) to calculate branch e i The wind resistance of the branch after the fault occurs, then the wind resistance matrix R' composed of the wind resistance of all branches of the mine ventilation network after the fault occurs is expressed as R' = {r'1, r'2, ..., r' n }, r′ i =r i ±Δr i ,(i=1,2,...,n) (6) In the formula, r i ' represents branch e i Branch windage after failure, r i Indicates branch e i The original wind resistance, Δr i Indicates branch e i The number of faults that occurred, n represents the number of all branches of the mine ventilation network; Step 2.4: The air volume of the mine ventilation network after the failure is solved by network solution. The air volume matrix Q' composed of the air volumes of all branches of the mine ventilation network after the failure is expressed as Q' = {q'1, q'2, ..., q' n }, the network solution is expressed as: Q'=f(R') (7) In the formula, Q' represents the air volume of the mine ventilation network after the fault, R' represents the wind resistance of the mine ventilation network after the fault, and f(*) represents the function of wind resistance and air volume; Step 2.5: The calculated wind resistance matrix R', air volume matrix Q' of the mine ventilation network after the failure, and the fault branch number e at each failure i , Fault quantity Δr i Recorded in the fault sample set.

3. According to claim 1, a method for diagnosing resistance-variable faults in a mine ventilation system based on a Bayesian optimized deep Q network is characterized in that: The step 3 is specifically described as follows: Step 3.1: Markov decision process for fault diagnosis of mine ventilation system Randomly select l samples D = {(Q1,e1),(Q2,e2),…,(Q i ,e i )..,(Q l ,e l )} constitutes l guessing questions, and then the agent guesses in the order of samples in D, where Q i is the air volume data corresponding to the sample, e i is the corresponding label, and finally the resistance-variable fault diagnosis problem of the mine ventilation system is transformed into a classification Markov decision process (CMDP), which is expressed as: CMDP = {S, A, R, P}; Where S represents the environmental state set, which depends on the training sample. In this study, it is the air volume set of each branch of the mine after the resistance type fault occurs. The first training sample x1 = {q 11 ,q 12 ,q 13 ,...,q 1n } as the initial state s1, state s t Corresponding sample x t , n is the number of branches; A represents the action set, that is, the set of fault branch numbers A = {e1, e2, e3, ..., e n }; R is the reward function, which comes from the feedback of the environment and is used to measure the pros and cons of the agent's execution of a certain action and guide the agent to learn the best strategy; P is the state transition probability matrix, which represents the rule p(s) for transferring from the current state to the next state after executing the action. t+1 s t ,a t ), in this study, P is determined and transferred according to the order of samples in the training sample. Step 3.2: WAFs-BODQN model structure design The WAFs-BODQN model contains two networks with the same structure but different parameters: Eval-Net and Target-Net. Eval-Net is used to update the current Q value, and Target-Net is used to calculate the fixed target Q value. One-dimensional convolutional neural network (1DCNN) is used as Eval-Net and Target-Net of WAFs-BODQN, and BO is used to optimize the parameters of 1DCNN. The whole process is expressed as follows: In the formula, σ1(·) is the activation function used in the output layer, pool represents the pooling operation, σ2(·) represents the activation function used in the activation layer, and M i is the set of feature maps output by the l-1th layer, is the jth output feature map of the l-1th layer, * is the convolution operation, is the i-th convolution kernel of the l-th layer, b l is the i-th bias of the l-th layer; BO is used to optimize the hyperparameter θ of 1DCNN, and the probability distribution of the optimization process is established using the Bayesian formula, which is expressed as formula (8): In the formula, p(y|G 1:i ) is the posterior probability distribution of y, p(G 1:i |y) is the likelihood distribution of y, p(G 1:i ) represents the marginal likelihood distribution, p(y) is the prior probability distribution of y, and y i =g(θ i )+ε i is the objective function value, g(·) is the objective function, ε i is the observation noise, G 1:i is the set of historical observations, θ i is the decision vector, i.e., the hyperparameter of CNN. Step 3.3: Reward function design Based on the KMeans++ clustering algorithm and the spatial distance function, the reward function is quantified by the Euclidean distance between samples. First, a sample is randomly selected as the cluster center u1, and then each sample x is calculated. i Minimum distance to cluster center u1: D(x i )=argmin||x i -u1||2 (10) Select the sample point with the largest distance as the next cluster center u2, repeat the above operation to finally get k clusters. Then, calculate the Euclidean distance between the samples in the same cluster and the cluster center and perform normalization: D(s i )=||x i -u j ||2,j=1,2,...,k (11) Set the reward value of the sample corresponding to the center of the majority class cluster to 1, and the reward value of the remaining samples to 1. As a bias, the reward value of the sample corresponding to the center of the minority cluster is set to the imbalance ratio IR, and the remaining samples are As a bias, the reward function is: The reward function R3 fully considers the imbalance and diversity of samples with the help of class spacing and intra-class spacing, enabling the agent to achieve better learning results. Step 3.4: Test the resistance-variable fault diagnosis method of the mine ventilation system based on the Bayesian optimized deep Q network constructed for the unbalanced data set, and verify the accuracy of the model diagnosis by comparing whether the fault branch number predicted by the model is the same as the actual fault branch number in the fault sample set.

Images